From 4f6e64e65c78f54d57ea2bdfba14b2b7259a2e27 Mon Sep 17 00:00:00 2001
From: sora <2075279110@qq.com>
Date: Thu, 17 Sep 2026 02:49:57 +0000
Subject: [PATCH] Stream docker pull progress; never rmi after a failed retag

Two pull UX/correctness fixes: (1) a multi-GB pull with captured
output is minutes of silence reading as a hang -- layer progress now
streams (throttled) to stderr; (2) the mirror-tag -> canonical retag
could fail silently and the following rmi then deleted the ONLY tag,
losing a 342s pull and forcing a full re-download -- retag is now
verified and the mirror tag kept on failure.

Co-Authored-By: Claude <noreply@anthropic.com>
---
 evalharness/sandbox/prefetch.py | 35 +++++++++++++++++++++++++++++----
 1 file changed, 31 insertions(+), 4 deletions(-)

diff --git a/evalharness/sandbox/prefetch.py b/evalharness/sandbox/prefetch.py
index f0d31c3..aa2b081 100644
--- a/evalharness/sandbox/prefetch.py
+++ b/evalharness/sandbox/prefetch.py
@@ -109,14 +109,41 @@ def _pull_one(image: str, verbose: bool = True) -> None:
             print(f'· pulling sandbox image {image} via {src} ...',
                   file=_sys.stderr, flush=True)
         t0 = _time.monotonic()
-        r = subprocess.run(['docker', 'pull', ref], capture_output=True, text=True,
-                           timeout=3600)
+        # stream progress: a multi-GB pull with captured output is minutes
+        # of silence that reads as a hang. Non-TTY docker pull emits clean
+        # per-layer lines; forward the informative ones (throttled).
+        proc = subprocess.Popen(['docker', 'pull', ref],
+                                stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
+                                text=True, bufsize=1)
+        last_line = ''
+        last_show = 0.0
+        import os as _os
+        for line in proc.stdout:
+            last_line = line.strip()
+            now = _time.monotonic()
+            if verbose and (any(k in line for k in
+                                ('Pulling from', 'Pull complete', 'Status',
+                                 'error', 'denied'))
+                            or now - last_show > 10):
+                print(f'  {last_line[:100]}', file=_sys.stderr, flush=True)
+                last_show = now
+        proc.wait(timeout=3600)
+        r = subprocess.CompletedProcess(proc.args, proc.returncode,
+                                        stdout='', stderr=last_line)
         if r.returncode == 0 and verbose:
             print(f'✓ image ready in {_time.monotonic() - t0:.0f}s',
                   file=_sys.stderr, flush=True)
         if r.returncode == 0:
-            if ref != image:  # retag the mirrored pull to the canonical name
-                subprocess.run(['docker', 'tag', ref, image], check=False)
+            if ref != image:
+                # retag BEFORE rmi and VERIFY: a silent tag failure followed
+                # by rmi deleted the only tag and lost a 342s multi-GB pull
+                tag = subprocess.run(['docker', 'tag', ref, image],
+                                     capture_output=True, text=True)
+                if tag.returncode != 0:
+                    print(f'!! docker tag {ref} -> {image} failed: '
+                          f'{(tag.stderr or "")[:120]} (keeping the mirror tag)',
+                          file=_sys.stderr, flush=True)
+                    return   # do NOT rmi: the mirror tag is the only handle
                 subprocess.run(['docker', 'rmi', ref], check=False)
             return
         last_err = (ref, (r.stderr or '').strip()[:150])